Financial text frequently contains cognitive bias triggers, such as an authority cue like “according to Goldman Sachs”, that can distort sentiment judgments without altering the underlying facts. While prior work has established that LLMs are susceptible to cognitive biases, existing evaluations treat bias exclusively as a property of the model, overlooking triggers already present in the input text. We ask whether LLMs can (i) detect such triggers, (ii) rewrite text to remove them, and (iii) whether doing so improves sentiment predictions, a task central to automated financial decision-making. We introduce TriggerBank, a dataset of 3,006 financial sentences with six types of synthetically injected bias triggers, and evaluate 11 LLMs across tasks (i)–(iii). Detection remains challenging, with framing and overconfidence hardest to detect, and removal fails for anchoring. Despite this, rewriting outperforms all instruction-based defenses for most models in downstream sentiment classification, while chain-of-thought prompting generally degrades performance. Debiasing the input text via rewriting is thus a more robust strategy than prompt-based approaches.